A full conference season, stats and video unified

The coaching platform answers a question a staff has during a game. This project answers a different one: what becomes possible when a full conference season of play-by-play, lineup data, and broadcast film lives in one system instead of three. One conference, one season, every game.

It is a fork of the real-time coaching platform pointed at an archive instead of a live feed: the 2025-26 Atlantic Coast Conference men’s season, all 429 games across 18 teams.

A season as a dataset

The ingestion side scrapes the conference’s public schedule and box scores, then pushes every game through the same pipeline the live platform uses: play-by-play normalized to one event stream, possession boundaries marked, five-man lineups tracked with automatic recovery of messy substitution data, and aggregated stats validated against the official box score. Scrape logs make the whole thing idempotent, so re-running a season cannot double-import a game.

Having the season in one schema is what makes the conference-level views cheap: standings with scoring margin, leader boards, and team pages are just queries.

The season dashboard: 18 teams, 429 games, 162 conference games, with standings led by Duke and a scoring leaders board. — The season at a glance: standings, scoring leaders, and 429 completed games.

Game pages that go past the box score

Each game carries the box score, lineup combinations with plus/minus, timeout and foul tracking, a filterable shot chart, and a win probability curve computed from the play-by-play. The screenshots on this page are from North Carolina 71, Duke 68, played February 7, 2026: 115 shots on the chart, 364 probability points on the curve.

The win probability chart for North Carolina against Duke: North Carolina favored for most of the game, with a sharp final swing at the end of the second half. — Win probability for the same game. North Carolina held the edge most of the way; the last minute is the spike.

The video layer

300 of the 429 games are joined to full-game broadcast recordings stored on my NAS, with the game clock mapped to video time. That is the piece that changes how the analytics get used: a lineup stint or a play-by-play row is not just a number, it is a timestamped pointer into film of the actual possessions.

It is also why this project stays on my hardware. The recordings are broadcast footage, roughly a terabyte of it, and it is not mine to republish. So there is no public demo of this one; it gets shown live, by screen share, which is how a film-plus-analytics tool gets used in a basketball office anyway.